Add REM confidence gating to prevent false positive dream playback

- Add remConfidence field to ClassificationResult3, SourceClassification, HybridClassification
- Compute remConfidence based on: REM probability, time >70min, REM window, consecutive signals
- Gate REM callbacks on MIN_REM_CONFIDENCE_FOR_PLAYBACK (0.5)
- Log skipped low-confidence REM detections for debugging
- Remove personal Fitbit data from git history and add to .gitignore

Author: jeremymanning

.gitignore

@@ -279,3 +279,4 @@ public/audio/*/*.mp3
 credentials.json
 notes/screenshots/*.png
 hardcoded_findings.txt
+notes/fitbit_merged_training_data.json

notes/classifier_deep_dive_20260115.md

@@ -0,0 +1,172 @@
+# Sleep Classifier Deep Dive - January 15, 2026
+
+## Executive Summary
+
+Analyzed 1,546 nights of Fitbit Takeout data with ~184k HR+stage samples to identify improvements for the sleep classifier. Key findings and recommendations below.
+
+## Current Performance
+
+- REM Sensitivity: 81.7%
+- Awake Sensitivity: 15.5%
+- Overall Accuracy: 41.9%
+
+## Data Analysis Findings
+
+### 1. Class Imbalance (CRITICAL)
+
+Fitbit labels show massive imbalance:
+
+- NREM: 68.6%
+- REM: 26.2%
+- **Wake: 5.3%** (only!)
+
+Wake samples concentrated at:
+
+- 30-60 min (sleep onset): 233 samples
+- 270-300 min (natural awakening): 293 samples
+
+**Implication**: Low awake precision is expected due to class imbalance. A 15% sensitivity with high specificity is actually reasonable.
+
+### 2. Feature Discrimination Analysis
+
+| Feature      | Wake     | Light | Deep | REM  | Best For                  |
+| ------------ | -------- | ----- | ---- | ---- | ------------------------- |
+| mean_hr      | 57.7     | 61.1  | 52.6 | 56.2 | REM vs NREM (+4.5 diff)   |
+| mean_diff    | 0.80     | 1.00  | 0.52 | 0.60 | REM vs NREM (+0.38)       |
+| **max_diff** | **4.13** | 3.44  | 1.84 | 2.07 | **Wake vs Sleep (+1.11)** |
+| hr_range     | 7.8      | 11.1  | 4.8  | 5.5  | REM vs NREM (+5.37)       |
+| std_hr       | 2.32     | 3.28  | 1.31 | 1.55 | REM vs NREM (+1.67)       |
+| rmssd        | 1.32     | 1.40  | 0.81 | 0.89 | REM vs NREM (+0.49)       |
+
+**Key Finding**: `max_diff` is the best discriminator for wake (4.13 vs 3.02 for sleep).
+
+### 3. Transition Probabilities (from 1,546 nights)
+
+```
+wake -> wake: 50.0%
+wake -> light: 28.9%
+wake -> rem: 11.5%
+wake -> deep: 9.6%
+
+deep -> light: 45.1%
+deep -> wake: 52.9%
+deep -> rem: 2.0%  (rare!)
+
+rem -> light: 45.8%
+rem -> wake: 52.8%
+rem -> deep: 1.4%  (rare!)
+```
+
+**Usable constraints**:
+
+- Deep rarely transitions directly to REM (2%)
+- REM rarely transitions to deep (1.4%)
+- Wake tends to stay wake (50%) or go to light (29%)
+
+### 4. Threshold Sweep Results
+
+For `max_diff` only:
+| Threshold | Wake Sens | Wake Prec | Sleep Spec |
+|-----------|-----------|-----------|------------|
+| 3.0 | 31.4% | 8.2% | 80.6% |
+| 4.0 | 23.0% | 10.6% | 89.2% |
+| 5.0 | 17.9% | 15.0% | 94.4% |
+
+Trade-off: Higher threshold = lower sensitivity but much better precision.
+
+## Recommendations
+
+### Short-term (High Impact, Low Effort)
+
+1. **Add `max_diff` as awake feature** alongside `mean_diff`
+   - Use threshold ~4.0 for high specificity
+   - Expected: +5-10% awake sensitivity
+
+2. **Add transition constraints**
+   - Block deep->rem and rem->deep transitions
+   - Expected: Fewer false REM classifications
+
+3. **Time-bin specific thresholds**
+   - Lower awake threshold at 30-60min (sleep onset)
+   - Lower awake threshold at 270-330min (natural wake time)
+
+### Medium-term (Moderate Effort)
+
+4. **Use `hr_range` and `std_hr` for REM discrimination**
+   - REM has distinctly LOWER hr_range (5.5 vs 10.8)
+   - REM has LOWER std_hr (1.55 vs 3.21)
+
+5. **Explore SpO2 data for REM**
+   - REM has irregular breathing → SpO2 variability
+   - 1,882 files available in Takeout
+
+6. **Ensemble approach**
+   - Combine current CV-based REM detector with new max_diff awake detector
+
+### Long-term (High Effort)
+
+7. **Train on full Fitbit Takeout**
+   - 104 sleep files, 6128 HR files available
+   - Potential for user-specific model fine-tuning
+
+8. **LSTM/Sequential model**
+   - Capture temporal patterns
+   - Literature suggests +10-15% improvement
+
+9. **Multi-modal features**
+   - Temperature changes correlate with sleep cycles
+   - Respiratory rate variability during REM
+
+## Available Data Summary
+
+| Source           | Files  | Resolution | Date Range |
+| ---------------- | ------ | ---------- | ---------- |
+| Sleep stages     | 104    | Per-stage  | 2017-2025  |
+| Heart rate       | 6,128  | 5 seconds  | 2017-2025  |
+| HRV (RMSSD)      | 1,869  | Daily      | 2020-2025  |
+| Respiratory rate | ~1,800 | Daily      | 2020-2025  |
+| SpO2             | 1,882  | 1 minute   | 2020-2025  |
+| Temperature      | 3,671  | Monthly    | 2020-2025  |
+
+## Next Steps
+
+1. ~~Implement `max_diff` feature in TypeScript classifier~~ ✅ DONE
+2. ~~Add transition probability constraints~~ (partial - time-specific thresholds added)
+3. ~~Test with time-specific thresholds~~ ✅ DONE
+4. Evaluate SpO2 integration feasibility
+
+## Session Update - January 15, 2026 (continued)
+
+### Implemented Changes
+
+1. **`remConfidence` field added to HybridClassification**
+   - Added to `SourceClassification` and `HybridClassification` interfaces
+   - Computed in `remOptimizedClassifier.ts` based on: stage probability, time >70min, REM window position, consecutive signals, REM propensity
+
+2. **REM playback gating on confidence**
+   - Added `MIN_REM_CONFIDENCE_FOR_PLAYBACK = 0.5` constant in `sleep.ts`
+   - Modified `handleStageTransition()` to check `remConfidence >= 0.5` before triggering REM callbacks
+   - Logs skipped REM detections with confidence level
+
+3. **Enhanced classifier features** (from earlier session)
+   - Added `max_diff` for awake detection
+   - Added `hr_range` for REM discrimination
+   - Added time-specific thresholds (lenient at 30-60min, 270-330min)
+
+### Build Output
+
+- APK: `~/Desktop/dream-stream-release-remconf.apk` (85.7 MB)
+- Build: Release variant
+- Ready for overnight testing
+
+### Expected Behavior
+
+- REM detections with confidence < 0.5 will be logged but not trigger dream playback
+- Should significantly reduce false positive REM detections while user is awake
+- Console log format: `[Sleep] REM detected but confidence X.XX < 0.5, skipping playback`
+
+## Files Created
+
+- `/scripts/parse_fitbit_takeout.py` - Data parsing and analysis
+- `/scripts/enhanced_classifier_v2.py` - New classifier prototype
+- `/notes/fitbit_merged_training_data.json` - 10k merged samples

services/hybridClassifier.ts

@@ -20,6 +20,7 @@ export interface StageProbabilities {
 export interface SourceClassification {
   probabilities: StageProbabilities;
   confidence: number;
+  remConfidence: number;
   available: boolean;
 }
 
@@ -29,6 +30,7 @@ export interface HybridClassification {
   fused: StageProbabilities;
   predictedStage: SleepStage;
   overallConfidence: number;
+  remConfidence: number;
 }
 
 type ClassifiableStage = 'awake' | 'light' | 'deep' | 'rem';
@@ -87,6 +89,7 @@ function classifyFromAudio(analysis: BreathingAnalysis | null): SourceClassifica
     return {
       probabilities: { awake: 0.25, light: 0.25, deep: 0.25, rem: 0.25 },
       confidence: 0,
+      remConfidence: 0,
       available: false,
     };
   }
@@ -143,6 +146,7 @@ function classifyFromAudio(analysis: BreathingAnalysis | null): SourceClassifica
   return {
     probabilities: normalizeProbabilities(probs),
     confidence: Math.min(1, confidence),
+    remConfidence: 0,
     available: true,
   };
 }
@@ -152,6 +156,7 @@ function classifyFromVitals(vitals: VitalsSnapshot | null): SourceClassification
     return {
       probabilities: { awake: 0.25, light: 0.25, deep: 0.25, rem: 0.25 },
       confidence: 0,
+      remConfidence: 0,
       available: false,
     };
   }
@@ -170,6 +175,7 @@ function classifyFromVitals(vitals: VitalsSnapshot | null): SourceClassification
   return {
     probabilities: normalizeProbabilities(probs),
     confidence: result.confidence,
+    remConfidence: result.remConfidence,
     available: true,
   };
 }
@@ -256,6 +262,7 @@ export async function classifyHybrid(
     fused,
     predictedStage,
     overallConfidence,
+    remConfidence: vitalsResult.remConfidence,
   };
 }
 

services/remOptimizedClassifier.ts

@@ -76,6 +76,7 @@ export interface ClassificationResult3 {
   stage: SleepStage3;
   stage4: SleepStage; // Mapped back to 4-class for compatibility
   confidence: number;
+  remConfidence: number;
   probabilities: Stage3Probabilities;
   temporalFeatures: TemporalFeatures;
   dataSource: 'vitals' | 'audio' | 'prediction' | 'hybrid';
@@ -1398,10 +1399,25 @@ export function classifyRemOptimized(
   previousStage = stage;
   previousProbabilities = probabilities;
 
+  // Compute REM-specific confidence for playback gating
+  let remConfidence = 0;
+  if (stage === 'rem') {
+    const baseRemConf = probabilities.rem;
+    const timeBonus = temporal.minutesSinceSleepStart > 70 ? 0.15 : 0;
+    const windowBonus = temporal.isInRemWindow ? 0.1 : 0;
+    const consecutiveBonus = Math.min(0.2, consecutiveRemSignals * 0.05);
+    const propensityBonus = temporal.remPropensity * 0.1;
+    remConfidence = Math.min(
+      1.0,
+      baseRemConf + timeBonus + windowBonus + consecutiveBonus + propensityBonus
+    );
+  }
+
   return {
     stage,
     stage4: to4Class(stage),
     confidence,
+    remConfidence,
     probabilities,
     temporalFeatures: temporal,
     dataSource,

services/sleep.ts

@@ -828,15 +828,27 @@ function inferSleepStage(analysis: BreathingAnalysis): SleepStage {
   return 'light';
 }
 
-function handleStageTransition(previousStage: SleepStage, newStage: SleepStage): void {
+const MIN_REM_CONFIDENCE_FOR_PLAYBACK = 0.5;
+
+function handleStageTransition(
+  previousStage: SleepStage,
+  newStage: SleepStage,
+  remConfidence: number
+): void {
   const now = Date.now();
   stageHistory.push({ stage: newStage, timestamp: now });
   stageHistory = stageHistory.filter((s) => now - s.timestamp < 3600000);
 
   notifyStageHistoryChange();
 
   if (newStage === 'rem' && previousStage !== 'rem') {
-    remCallbacks.forEach((cb) => cb());
+    if (remConfidence >= MIN_REM_CONFIDENCE_FOR_PLAYBACK) {
+      remCallbacks.forEach((cb) => cb());
+    } else {
+      console.log(
+        `[Sleep] REM detected but confidence ${remConfidence.toFixed(2)} < ${MIN_REM_CONFIDENCE_FOR_PLAYBACK}, skipping playback`
+      );
+    }
   }
 
   if (previousStage === 'rem' && newStage !== 'rem') {
@@ -853,7 +865,7 @@ async function updateStageFromAudio(analysis: BreathingAnalysis): Promise<void>
   if (hybrid.predictedStage !== currentSession.currentStage) {
     const previousStage = currentSession.currentStage;
     updateSleepStage(hybrid.predictedStage);
-    handleStageTransition(previousStage, hybrid.predictedStage);
+    handleStageTransition(previousStage, hybrid.predictedStage, hybrid.remConfidence);
   }
 }
 
@@ -988,7 +1000,7 @@ export async function processVitalsUpdate(vitals: VitalsSnapshot): Promise<void>
       if (hybrid.predictedStage !== currentSession.currentStage) {
         const previousStage = currentSession.currentStage;
         updateSleepStage(hybrid.predictedStage);
-        handleStageTransition(previousStage, hybrid.predictedStage);
+        handleStageTransition(previousStage, hybrid.predictedStage, hybrid.remConfidence);
       }
     }
   }